Vast.ai

Vast.ai prices per GPU per hour and sizes every requirement per card. A machine with 8 cards must clear 8 times the RAM, disk and port thresholds of a machine with 1.

Connecting

1. Get an API key from your Vast.ai account.

2. Add the account in Flotera under Marketplaces. The key is stored encrypted and never shown again.

   Listings and rentals that already exist on Vast appear in the panel within a minute or two. Connecting does not change anything on their side.

3. Install their agent from the machine’s Platforms tab.

   This is Vast’s own daemon, not ours. It is what actually rents the machine out — until it is installed and has checked in, everything else is preparation.

4. Run the readiness check and fix what it lists.

5. Set a price and go live.

Verification

Vast verifies machines before they reach the full market, and verification checks things the readiness check also covers — AVX, CUDA, the per-GPU minimums. Fixing the readiness rows first is the cheaper order: a machine that fails verification has to go round again.

Two rows are worth checking before you buy hardware, because neither is fixable afterwards:

• AVX support — a property of the processor.
• VRAM per GPU — fixed at manufacture, and below 7 GB the card is filtered out of every modern template.

Pricing

Vast prices per GPU-hour. The panel shows your price against the market for the same card — the median, and where you sit relative to it — on the same screen where you set it.

The market number is the useful one

A price is not high or low on its own; it is high or low against what identical cards are renting for right now, and how many of them are idle. Both figures are on the pricing screen.

Changing a price shows what the change means before you commit it. Nothing is sent to Vast until you confirm.

Going in and out of search

Show in search publishes the machine to renters. Hide from search takes it out without deleting the listing — the offer survives, renters just stop seeing it.

Hiding is the right move before maintenance, and it is what the guards ask you to do before a reboot.

Listed is not the same as rented

A machine in search earns nothing until someone books it. If it stays in search and idle for a long time, the price is usually the reason — compare it against the median on the pricing screen.

CUDA is a market-reach decision

Vast publishes what each CUDA step costs you in reachable workloads:

CUDA version	Share of workloads that can use the machine
11	about 30%
12	about 85%
13	all of them

CUDA comes with the driver, so this is a driver decision. Update while the machine is idle: the update wants a reboot, and a reboot under a tenant breaks the contract.

Clore.aiA different pricing model, different fees, and currencies to choose.